Answer
What belongs in the main agent versus a worker agent?
Decisions and actions stay in the main run. Reading and searching go to workers. The split is by consequence, not difficulty.
The main run keeps the decisions, the accumulated state and anything that changes something. Workers take the reading, searching and summarising. The division is by consequence rather than by how hard the task is.
The main run is the only participant that holds the whole situation, which makes it the only one positioned to decide anything. A worker sees a brief and its own findings; it does not know what else is true, what has already been settled, or what the decision interacts with. Decisions made there are made on a partial view and arrive as a line in a summary, which is the worst possible place for anyone to notice them.
So the division follows consequence. Anything that changes state, commits the business, or forecloses an option belongs in the main run where the full context sits and where the action is visible. Anything that produces knowledge — where is this, does this exist, what does this document say about that, which of these match — is a worker's job, and the more it has to read to answer, the better the fit.
The main run also holds the state, and this is worth stating separately because it is where the arrangement most often goes wrong. Workers should not be maintaining the plan, updating the record of decisions, or writing to the shared state file. If two workers do that concurrently the state becomes a merge problem, and if one does it the main run's picture is being edited by a participant that cannot see the whole picture.
Difficulty is a misleading axis and is the one people reach for. Delegating the hard part sounds sensible and produces the opposite of what is wanted: the hard part is usually where the judgement is, and judgement performed with less context is worse. Delegating the tedious part — the wide reading, the enumeration, the cross-referencing — is what actually helps, and it is the part that felt too menial to hand over.
There is a boundary case that is worth deciding deliberately: a worker that must act to find something out, such as running a query or executing a command to observe a result. That is reading in intent and acting in mechanism, and the resolution is usually to permit it where the action is read-only and reversible, and to have the worker report rather than act where it is not. What matters is that the rule is set rather than left to the worker's assessment.
The test that catches most misplacements: if the worker's output would be accepted without anyone looking at it, whatever it did is effectively a decision. Either the output needs to be reviewed or the work should not have been delegated, and noticing which of those applies is the useful question about any proposed division.
Send out the reading and keep the deciding, because a decision made in a worker is a decision nobody reviewed.
Siddharth Sharma, Context Theory
Related questions
Should the main run do any reading at all?
Yes, for anything it needs to reason about directly. Delegating every read means the main run holds only summaries, and a decision made entirely from summaries is a decision made from a lossy account of the situation. Delegate the wide reading that produces a fact; keep the reading of the things the decision turns on.
Can a worker spawn its own workers?
Technically usually, and it degrades traceability sharply, because the chain of what was asked and what was returned now has a layer nobody designed. Where a worker's task is large enough to need dividing, that is a signal the division belongs in the main run, which is the participant that can see whether the pieces cover the work.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
2026 speed-to-lead benchmark · verified
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | The main run is the only participant holding the full situation, so a decision made in a worker is made on a partial view and surfaces as a line in a summary, which is the least visible place for it to be noticed. | Checking a worker's returned summary for choices it made that were not in its brief. |
| Software | Shared state must be written only by the main run, because concurrent worker writes create a merge problem and a single worker write allows a participant without the full picture to edit the main run's model of the situation. | Checking whether any delegated run has write access to the plan or decisions file. |
| Response | Delegating by difficulty misplaces judgement into a participant with less context, whereas delegating by tedium moves wide reading and enumeration, which is where the context saving actually comes from. | Classifying each candidate delegation by whether its output is a judgement or a fact. |
| Constraint | If a worker's output would be accepted without inspection, the work it performed was effectively a decision, which makes the acceptance test the practical check on any proposed division. | Asking whether the returned result would in practice be read before being acted on. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one